panfrost: add atomic_cmpxchg opcode
[mesa.git] / src / panfrost / midgard / midgard.h
1 /* Author(s):
2 * Connor Abbott
3 * Alyssa Rosenzweig
4 *
5 * Copyright (c) 2013 Connor Abbott (connor@abbott.cx)
6 * Copyright (c) 2018 Alyssa Rosenzweig (alyssa@rosenzweig.io)
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a copy
9 * of this software and associated documentation files (the "Software"), to deal
10 * in the Software without restriction, including without limitation the rights
11 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
12 * copies of the Software, and to permit persons to whom the Software is
13 * furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice shall be included in
16 * all copies or substantial portions of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
21 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
24 * THE SOFTWARE.
25 */
26
27 #ifndef __midgard_h__
28 #define __midgard_h__
29
30 #include <stdint.h>
31 #include <stdbool.h>
32
33 #define MIDGARD_DBG_MSGS 0x0001
34 #define MIDGARD_DBG_SHADERS 0x0002
35 #define MIDGARD_DBG_SHADERDB 0x0004
36
37 extern int midgard_debug;
38
39 typedef enum {
40 midgard_word_type_alu,
41 midgard_word_type_load_store,
42 midgard_word_type_texture,
43 midgard_word_type_unknown
44 } midgard_word_type;
45
46 typedef enum {
47 midgard_alu_vmul,
48 midgard_alu_sadd,
49 midgard_alu_smul,
50 midgard_alu_vadd,
51 midgard_alu_lut
52 } midgard_alu;
53
54 enum {
55 TAG_INVALID = 0x0,
56 TAG_BREAK = 0x1,
57 TAG_TEXTURE_4_VTX = 0x2,
58 TAG_TEXTURE_4 = 0x3,
59 TAG_TEXTURE_4_BARRIER = 0x4,
60 TAG_LOAD_STORE_4 = 0x5,
61 TAG_UNKNOWN_1 = 0x6,
62 TAG_UNKNOWN_2 = 0x7,
63 TAG_ALU_4 = 0x8,
64 TAG_ALU_8 = 0x9,
65 TAG_ALU_12 = 0xA,
66 TAG_ALU_16 = 0xB,
67 TAG_ALU_4_WRITEOUT = 0xC,
68 TAG_ALU_8_WRITEOUT = 0xD,
69 TAG_ALU_12_WRITEOUT = 0xE,
70 TAG_ALU_16_WRITEOUT = 0xF
71 };
72
73 /*
74 * ALU words
75 */
76
77 typedef enum {
78 midgard_alu_op_fadd = 0x10,
79 midgard_alu_op_fmul = 0x14,
80
81 midgard_alu_op_fmin = 0x28,
82 midgard_alu_op_fmax = 0x2C,
83
84 midgard_alu_op_fmov = 0x30, /* fmov_rte */
85 midgard_alu_op_fmov_rtz = 0x31,
86 midgard_alu_op_fmov_rtn = 0x32,
87 midgard_alu_op_fmov_rtp = 0x33,
88 midgard_alu_op_froundeven = 0x34,
89 midgard_alu_op_ftrunc = 0x35,
90 midgard_alu_op_ffloor = 0x36,
91 midgard_alu_op_fceil = 0x37,
92 midgard_alu_op_ffma = 0x38,
93 midgard_alu_op_fdot3 = 0x3C,
94 midgard_alu_op_fdot3r = 0x3D,
95 midgard_alu_op_fdot4 = 0x3E,
96 midgard_alu_op_freduce = 0x3F,
97
98 midgard_alu_op_iadd = 0x40,
99 midgard_alu_op_ishladd = 0x41, /* a + (b<<1) */
100 midgard_alu_op_isub = 0x46,
101 midgard_alu_op_iaddsat = 0x48,
102 midgard_alu_op_uaddsat = 0x49,
103 midgard_alu_op_isubsat = 0x4E,
104 midgard_alu_op_usubsat = 0x4F,
105
106 midgard_alu_op_imul = 0x58,
107
108 midgard_alu_op_imin = 0x60,
109 midgard_alu_op_umin = 0x61,
110 midgard_alu_op_imax = 0x62,
111 midgard_alu_op_umax = 0x63,
112 midgard_alu_op_ihadd = 0x64,
113 midgard_alu_op_uhadd = 0x65,
114 midgard_alu_op_irhadd = 0x66,
115 midgard_alu_op_urhadd = 0x67,
116 midgard_alu_op_iasr = 0x68,
117 midgard_alu_op_ilsr = 0x69,
118 midgard_alu_op_ishl = 0x6E,
119
120 midgard_alu_op_iand = 0x70,
121 midgard_alu_op_ior = 0x71,
122 midgard_alu_op_inand = 0x72, /* ~(a & b), for inot let a = b */
123 midgard_alu_op_inor = 0x73, /* ~(a | b) */
124 midgard_alu_op_iandnot = 0x74, /* (a & ~b), used for not/b2f */
125 midgard_alu_op_iornot = 0x75, /* (a | ~b) */
126 midgard_alu_op_ixor = 0x76,
127 midgard_alu_op_inxor = 0x77, /* ~(a & b) */
128 midgard_alu_op_iclz = 0x78, /* Number of zeroes on left */
129 midgard_alu_op_ibitcount8 = 0x7A, /* Counts bits in 8-bit increments */
130 midgard_alu_op_imov = 0x7B,
131 midgard_alu_op_iabsdiff = 0x7C,
132 midgard_alu_op_uabsdiff = 0x7D,
133 midgard_alu_op_ichoose = 0x7E, /* vector, component number - dupe for shuffle() */
134
135 midgard_alu_op_feq = 0x80,
136 midgard_alu_op_fne = 0x81,
137 midgard_alu_op_flt = 0x82,
138 midgard_alu_op_fle = 0x83,
139 midgard_alu_op_fball_eq = 0x88,
140 midgard_alu_op_fball_neq = 0x89,
141 midgard_alu_op_fball_lt = 0x8A, /* all(lessThan(.., ..)) */
142 midgard_alu_op_fball_lte = 0x8B, /* all(lessThanEqual(.., ..)) */
143
144 midgard_alu_op_fbany_eq = 0x90,
145 midgard_alu_op_fbany_neq = 0x91,
146 midgard_alu_op_fbany_lt = 0x92, /* any(lessThan(.., ..)) */
147 midgard_alu_op_fbany_lte = 0x93, /* any(lessThanEqual(.., ..)) */
148
149 midgard_alu_op_f2i_rte = 0x98,
150 midgard_alu_op_f2i_rtz = 0x99,
151 midgard_alu_op_f2i_rtn = 0x9A,
152 midgard_alu_op_f2i_rtp = 0x9B,
153 midgard_alu_op_f2u_rte = 0x9C,
154 midgard_alu_op_f2u_rtz = 0x9D,
155 midgard_alu_op_f2u_rtn = 0x9E,
156 midgard_alu_op_f2u_rtp = 0x9F,
157
158 midgard_alu_op_ieq = 0xA0,
159 midgard_alu_op_ine = 0xA1,
160 midgard_alu_op_ult = 0xA2,
161 midgard_alu_op_ule = 0xA3,
162 midgard_alu_op_ilt = 0xA4,
163 midgard_alu_op_ile = 0xA5,
164 midgard_alu_op_iball_eq = 0xA8,
165 midgard_alu_op_iball_neq = 0xA9,
166 midgard_alu_op_uball_lt = 0xAA,
167 midgard_alu_op_uball_lte = 0xAB,
168 midgard_alu_op_iball_lt = 0xAC,
169 midgard_alu_op_iball_lte = 0xAD,
170
171 midgard_alu_op_ibany_eq = 0xB0,
172 midgard_alu_op_ibany_neq = 0xB1,
173 midgard_alu_op_ubany_lt = 0xB2,
174 midgard_alu_op_ubany_lte = 0xB3,
175 midgard_alu_op_ibany_lt = 0xB4, /* any(lessThan(.., ..)) */
176 midgard_alu_op_ibany_lte = 0xB5, /* any(lessThanEqual(.., ..)) */
177 midgard_alu_op_i2f_rte = 0xB8,
178 midgard_alu_op_i2f_rtz = 0xB9,
179 midgard_alu_op_i2f_rtn = 0xBA,
180 midgard_alu_op_i2f_rtp = 0xBB,
181 midgard_alu_op_u2f_rte = 0xBC,
182 midgard_alu_op_u2f_rtz = 0xBD,
183 midgard_alu_op_u2f_rtn = 0xBE,
184 midgard_alu_op_u2f_rtp = 0xBF,
185
186 midgard_alu_op_icsel_v = 0xC0, /* condition code r31 */
187 midgard_alu_op_icsel = 0xC1, /* condition code r31.w */
188 midgard_alu_op_fcsel_v = 0xC4,
189 midgard_alu_op_fcsel = 0xC5,
190 midgard_alu_op_fround = 0xC6,
191
192 midgard_alu_op_fatan_pt2 = 0xE8,
193 midgard_alu_op_fpow_pt1 = 0xEC,
194 midgard_alu_op_fpown_pt1 = 0xED,
195 midgard_alu_op_fpowr_pt1 = 0xEE,
196
197 midgard_alu_op_frcp = 0xF0,
198 midgard_alu_op_frsqrt = 0xF2,
199 midgard_alu_op_fsqrt = 0xF3,
200 midgard_alu_op_fexp2 = 0xF4,
201 midgard_alu_op_flog2 = 0xF5,
202 midgard_alu_op_fsin = 0xF6,
203 midgard_alu_op_fcos = 0xF7,
204 midgard_alu_op_fatan2_pt1 = 0xF9,
205 } midgard_alu_op;
206
207 typedef enum {
208 midgard_outmod_none = 0,
209 midgard_outmod_pos = 1, /* max(x, 0.0) */
210 midgard_outmod_sat_signed = 2, /* clamp(x, -1.0, 1.0) */
211 midgard_outmod_sat = 3 /* clamp(x, 0.0, 1.0) */
212 } midgard_outmod_float;
213
214 typedef enum {
215 midgard_outmod_int_saturate = 0,
216 midgard_outmod_uint_saturate = 1,
217 midgard_outmod_int_wrap = 2,
218 midgard_outmod_int_high = 3, /* Overflowed portion */
219 } midgard_outmod_int;
220
221 typedef enum {
222 midgard_reg_mode_8 = 0,
223 midgard_reg_mode_16 = 1,
224 midgard_reg_mode_32 = 2,
225 midgard_reg_mode_64 = 3
226 } midgard_reg_mode;
227
228 typedef enum {
229 midgard_dest_override_lower = 0,
230 midgard_dest_override_upper = 1,
231 midgard_dest_override_none = 2
232 } midgard_dest_override;
233
234 typedef enum {
235 midgard_int_sign_extend = 0,
236 midgard_int_zero_extend = 1,
237 midgard_int_normal = 2,
238 midgard_int_shift = 3
239 } midgard_int_mod;
240
241 #define MIDGARD_FLOAT_MOD_ABS (1 << 0)
242 #define MIDGARD_FLOAT_MOD_NEG (1 << 1)
243
244 typedef struct
245 __attribute__((__packed__))
246 {
247 /* Either midgard_int_mod or from midgard_float_mod_*, depending on the
248 * type of op */
249 unsigned mod : 2;
250
251 /* replicate lower half if dest = half, or low/high half selection if
252 * dest = full
253 */
254 bool rep_low : 1;
255 bool rep_high : 1; /* unused if dest = full */
256 bool half : 1; /* only matters if dest = full */
257 unsigned swizzle : 8;
258 }
259 midgard_vector_alu_src;
260
261 typedef struct
262 __attribute__((__packed__))
263 {
264 midgard_alu_op op : 8;
265 midgard_reg_mode reg_mode : 2;
266 unsigned src1 : 13;
267 unsigned src2 : 13;
268 midgard_dest_override dest_override : 2;
269 unsigned outmod : 2;
270 unsigned mask : 8;
271 }
272 midgard_vector_alu;
273
274 typedef struct
275 __attribute__((__packed__))
276 {
277 unsigned mod : 2;
278 bool full : 1; /* 0 = half, 1 = full */
279 unsigned component : 3;
280 }
281 midgard_scalar_alu_src;
282
283 typedef struct
284 __attribute__((__packed__))
285 {
286 midgard_alu_op op : 8;
287 unsigned src1 : 6;
288 unsigned src2 : 11;
289 unsigned unknown : 1;
290 unsigned outmod : 2;
291 bool output_full : 1;
292 unsigned output_component : 3;
293 }
294 midgard_scalar_alu;
295
296 typedef struct
297 __attribute__((__packed__))
298 {
299 unsigned src1_reg : 5;
300 unsigned src2_reg : 5;
301 unsigned out_reg : 5;
302 bool src2_imm : 1;
303 }
304 midgard_reg_info;
305
306 /* In addition to conditional branches and jumps (unconditional branches),
307 * Midgard implements a bit of fixed function functionality used in fragment
308 * shaders via specially crafted branches. These have special branch opcodes,
309 * which perform a fixed-function operation and/or use the results of a
310 * fixed-function operation as the branch condition. */
311
312 typedef enum {
313 /* Regular branches */
314 midgard_jmp_writeout_op_branch_uncond = 1,
315 midgard_jmp_writeout_op_branch_cond = 2,
316
317 /* In a fragment shader, execute a discard_if instruction, with the
318 * corresponding condition code. Terminates the shader, so generally
319 * set the branch target to out of the shader */
320 midgard_jmp_writeout_op_discard = 4,
321
322 /* Branch if the tilebuffer is not yet ready. At the beginning of a
323 * fragment shader that reads from the tile buffer, for instance via
324 * ARM_shader_framebuffer_fetch or EXT_pixel_local_storage, this branch
325 * operation should be used as a loop. An instruction like
326 * "br.tilebuffer.always -1" does the trick, corresponding to
327 * "while(!is_tilebuffer_ready) */
328 midgard_jmp_writeout_op_tilebuffer_pending = 6,
329
330 /* In a fragment shader, try to write out the value pushed to r0 to the
331 * tilebuffer, subject to unknown state in r1.z and r1.w. If this
332 * succeeds, the shader terminates. If it fails, it branches to the
333 * specified branch target. Generally, this should be used in a loop to
334 * itself, acting as "do { write(r0); } while(!write_successful);" */
335 midgard_jmp_writeout_op_writeout = 7,
336 } midgard_jmp_writeout_op;
337
338 typedef enum {
339 midgard_condition_write0 = 0,
340
341 /* These condition codes denote a conditional branch on FALSE and on
342 * TRUE respectively */
343 midgard_condition_false = 1,
344 midgard_condition_true = 2,
345
346 /* This condition code always branches. For a pure branch, the
347 * unconditional branch coding should be used instead, but for
348 * fixed-function branch opcodes, this is still useful */
349 midgard_condition_always = 3,
350 } midgard_condition;
351
352 typedef struct
353 __attribute__((__packed__))
354 {
355 midgard_jmp_writeout_op op : 3; /* == branch_uncond */
356 unsigned dest_tag : 4; /* tag of branch destination */
357 unsigned unknown : 2;
358 int offset : 7;
359 }
360 midgard_branch_uncond;
361
362 typedef struct
363 __attribute__((__packed__))
364 {
365 midgard_jmp_writeout_op op : 3; /* == branch_cond */
366 unsigned dest_tag : 4; /* tag of branch destination */
367 int offset : 7;
368 midgard_condition cond : 2;
369 }
370 midgard_branch_cond;
371
372 typedef struct
373 __attribute__((__packed__))
374 {
375 midgard_jmp_writeout_op op : 3; /* == branch_cond */
376 unsigned dest_tag : 4; /* tag of branch destination */
377 unsigned unknown : 2;
378 signed offset : 23;
379
380 /* Extended branches permit inputting up to 4 conditions loaded into
381 * r31 (two in r31.w and two in r31.x). In the most general case, we
382 * specify a function f(A, B, C, D) mapping 4 1-bit conditions to a
383 * single 1-bit branch criteria. Note that the domain of f has 2^(2^4)
384 * elements, each mapping to 1-bit of output, so we can trivially
385 * construct a Godel numbering of f as a (2^4)=16-bit integer. This
386 * 16-bit integer serves as a lookup table to compute f, subject to
387 * some swaps for ordering.
388 *
389 * Interesting, the standard 2-bit condition codes are also a LUT with
390 * the same format (2^1-bit), but it's usually easier to use enums. */
391
392 unsigned cond : 16;
393 }
394 midgard_branch_extended;
395
396 typedef struct
397 __attribute__((__packed__))
398 {
399 midgard_jmp_writeout_op op : 3; /* == writeout */
400 unsigned unknown : 13;
401 }
402 midgard_writeout;
403
404 /*
405 * Load/store words
406 */
407
408 typedef enum {
409 midgard_op_ld_st_noop = 0x03,
410
411 /* Unpack a colour from a native format to fp16 */
412 midgard_op_unpack_colour = 0x05,
413
414 /* Packs a colour from fp16 to a native format */
415 midgard_op_pack_colour = 0x09,
416
417 /* Likewise packs from fp32 */
418 midgard_op_pack_colour_32 = 0x0A,
419
420 /* Unclear why this is on the L/S unit, but moves fp32 cube map
421 * coordinates in r27 to its cube map texture coordinate destination
422 * (e.g r29). */
423
424 midgard_op_ld_cubemap_coords = 0x0E,
425
426 /* Loads a global/local/group ID, depending on arguments */
427 midgard_op_ld_compute_id = 0x10,
428
429 /* The L/S unit can do perspective division a clock faster than the ALU
430 * if you're lucky. Put the vec4 in r27, and call with 0x24 as the
431 * unknown state; the output will be <x/w, y/w, z/w, 1>. Replace w with
432 * z for the z version */
433 midgard_op_ldst_perspective_division_z = 0x12,
434 midgard_op_ldst_perspective_division_w = 0x13,
435
436 /* val in r27.y, address embedded, outputs result to argument. Invert val for sub. Let val = +-1 for inc/dec. */
437 midgard_op_atomic_add = 0x40,
438 midgard_op_atomic_add64 = 0x41,
439
440 midgard_op_atomic_and = 0x44,
441 midgard_op_atomic_and64 = 0x45,
442 midgard_op_atomic_or = 0x48,
443 midgard_op_atomic_or64 = 0x49,
444 midgard_op_atomic_xor = 0x4C,
445 midgard_op_atomic_xor64 = 0x4D,
446
447 midgard_op_atomic_imin = 0x50,
448 midgard_op_atomic_imin64 = 0x51,
449 midgard_op_atomic_umin = 0x54,
450 midgard_op_atomic_umin64 = 0x55,
451 midgard_op_atomic_imax = 0x58,
452 midgard_op_atomic_imax64 = 0x59,
453 midgard_op_atomic_umax = 0x5C,
454 midgard_op_atomic_umax64 = 0x5D,
455
456 midgard_op_atomic_xchg = 0x60,
457 midgard_op_atomic_xchg64 = 0x61,
458
459 midgard_op_atomic_cmpxchg = 0x64,
460 midgard_op_atomic_cmpxchg64 = 0x65,
461
462 /* Used for compute shader's __global arguments, __local variables (or
463 * for register spilling) */
464
465 midgard_op_ld_uchar = 0x80, /* zero extends */
466 midgard_op_ld_char = 0x81, /* sign extends */
467 midgard_op_ld_ushort = 0x84, /* zero extends */
468 midgard_op_ld_short = 0x85, /* sign extends */
469 midgard_op_ld_char4 = 0x88, /* short2, int, float */
470 midgard_op_ld_short4 = 0x8C, /* int2, float2, long */
471 midgard_op_ld_int4 = 0x90, /* float4, long2 */
472
473 midgard_op_ld_attr_32 = 0x94,
474 midgard_op_ld_attr_16 = 0x95,
475 midgard_op_ld_attr_32u = 0x96,
476 midgard_op_ld_attr_32i = 0x97,
477 midgard_op_ld_vary_32 = 0x98,
478 midgard_op_ld_vary_16 = 0x99,
479 midgard_op_ld_vary_32u = 0x9A,
480 midgard_op_ld_vary_32i = 0x9B,
481
482 /* Old version of midgard_op_ld_color_buffer_as_fp16, for T720 */
483 midgard_op_ld_color_buffer_as_fp32_old = 0x9C,
484 midgard_op_ld_color_buffer_as_fp16_old = 0x9D,
485 midgard_op_ld_color_buffer_32u_old = 0x9E,
486
487 /* The distinction between these ops is the alignment requirement /
488 * accompanying shift. Thus, the offset to ld_ubo_int4 is in 16-byte
489 * units and can load 128-bit. The offset to ld_ubo_short4 is in 8-byte
490 * units; ld_ubo_char4 in 4-byte units. ld_ubo_char/ld_ubo_char2 are
491 * purely theoretical (never seen in the wild) since int8/int16/fp16
492 * UBOs don't really exist. The ops are still listed to maintain
493 * symmetry with generic I/O ops. */
494
495 midgard_op_ld_ubo_char = 0xA0, /* theoretical */
496 midgard_op_ld_ubo_char2 = 0xA4, /* theoretical */
497 midgard_op_ld_ubo_char4 = 0xA8,
498 midgard_op_ld_ubo_short4 = 0xAC,
499 midgard_op_ld_ubo_int4 = 0xB0,
500
501 /* New-style blending ops. Works on T760/T860 */
502 midgard_op_ld_color_buffer_as_fp32 = 0xB8,
503 midgard_op_ld_color_buffer_as_fp16 = 0xB9,
504 midgard_op_ld_color_buffer_32u = 0xBA,
505
506 midgard_op_st_char = 0xC0,
507 midgard_op_st_char2 = 0xC4, /* short */
508 midgard_op_st_char4 = 0xC8, /* short2, int, float */
509 midgard_op_st_short4 = 0xCC, /* int2, float2, long */
510 midgard_op_st_int4 = 0xD0, /* float4, long2 */
511
512 midgard_op_st_vary_32 = 0xD4,
513 midgard_op_st_vary_16 = 0xD5,
514 midgard_op_st_vary_32u = 0xD6,
515 midgard_op_st_vary_32i = 0xD7,
516
517 /* Value to st in r27, location r26.w as short2 */
518 midgard_op_st_image_f = 0xD8,
519 midgard_op_st_image_ui = 0xDA,
520 midgard_op_st_image_i = 0xDB,
521 } midgard_load_store_op;
522
523 typedef enum {
524 midgard_interp_sample = 0,
525 midgard_interp_centroid = 1,
526 midgard_interp_default = 2
527 } midgard_interpolation;
528
529 typedef enum {
530 midgard_varying_mod_none = 0,
531
532 /* Other values unknown */
533
534 /* Take the would-be result and divide all components by its z/w
535 * (perspective division baked in with the load) */
536 midgard_varying_mod_perspective_z = 2,
537 midgard_varying_mod_perspective_w = 3,
538 } midgard_varying_modifier;
539
540 typedef struct
541 __attribute__((__packed__))
542 {
543 unsigned zero0 : 1; /* Always zero */
544
545 midgard_varying_modifier modifier : 2;
546
547 unsigned zero1: 1; /* Always zero */
548
549 /* Varying qualifiers, zero if not a varying */
550 unsigned flat : 1;
551 unsigned is_varying : 1; /* Always one for varying, but maybe something else? */
552 midgard_interpolation interpolation : 2;
553
554 unsigned zero2 : 2; /* Always zero */
555 }
556 midgard_varying_parameter;
557
558 /* 8-bit register/etc selector for load/store ops */
559 typedef struct
560 __attribute__((__packed__))
561 {
562 /* Indexes into the register */
563 unsigned component : 2;
564
565 /* Register select between r26/r27 */
566 unsigned select : 1;
567
568 unsigned unknown : 2;
569
570 /* Like any good Arm instruction set, load/store arguments can be
571 * implicitly left-shifted... but only the second argument. Zero for no
572 * shifting, up to <<7 possible though. This is useful for indexing.
573 *
574 * For the first argument, it's unknown what these bits mean */
575 unsigned shift : 3;
576 }
577 midgard_ldst_register_select;
578
579 typedef struct
580 __attribute__((__packed__))
581 {
582 midgard_load_store_op op : 8;
583 unsigned reg : 5;
584 unsigned mask : 4;
585 unsigned swizzle : 8;
586
587 /* Load/store ops can take two additional registers as arguments, but
588 * these are limited to load/store registers with only a few supported
589 * mask/swizzle combinations. The tradeoff is these are much more
590 * compact, requiring 8-bits each rather than 17-bits for a full
591 * reg/mask/swizzle. Usually (?) encoded as
592 * midgard_ldst_register_select. */
593 unsigned arg_1 : 8;
594 unsigned arg_2 : 8;
595
596 unsigned varying_parameters : 10;
597
598 unsigned address : 9;
599 }
600 midgard_load_store_word;
601
602 typedef struct
603 __attribute__((__packed__))
604 {
605 unsigned type : 4;
606 unsigned next_type : 4;
607 uint64_t word1 : 60;
608 uint64_t word2 : 60;
609 }
610 midgard_load_store;
611
612 /* 8-bit register selector used in texture ops to select a bias/LOD/gradient
613 * register, shoved into the `bias` field */
614
615 typedef struct
616 __attribute__((__packed__))
617 {
618 /* 32-bit register, clear for half-register */
619 unsigned full : 1;
620
621 /* Register select between r28/r29 */
622 unsigned select : 1;
623
624 /* For a half-register, selects the upper half */
625 unsigned upper : 1;
626
627 /* Indexes into the register */
628 unsigned component : 2;
629
630 /* Padding to make this 8-bit */
631 unsigned zero : 3;
632 }
633 midgard_tex_register_select;
634
635 /* Texture pipeline results are in r28-r29 */
636 #define REG_TEX_BASE 28
637
638 enum mali_texture_op {
639 TEXTURE_OP_NORMAL = 1, /* texture */
640 TEXTURE_OP_LOD = 2, /* textureLod */
641 TEXTURE_OP_TEXEL_FETCH = 4,
642 TEXTURE_OP_BARRIER = 11,
643 TEXTURE_OP_DERIVATIVE = 13
644 };
645
646 enum mali_sampler_type {
647 MALI_SAMPLER_UNK = 0x0,
648 MALI_SAMPLER_FLOAT = 0x1, /* sampler */
649 MALI_SAMPLER_UNSIGNED = 0x2, /* usampler */
650 MALI_SAMPLER_SIGNED = 0x3, /* isampler */
651 };
652
653 /* Texture modes */
654 enum mali_texture_mode {
655 TEXTURE_NORMAL = 1,
656 TEXTURE_SHADOW = 5,
657 TEXTURE_GATHER_SHADOW = 6,
658 TEXTURE_GATHER_X = 8,
659 TEXTURE_GATHER_Y = 9,
660 TEXTURE_GATHER_Z = 10,
661 TEXTURE_GATHER_W = 11,
662 };
663
664 enum mali_derivative_mode {
665 TEXTURE_DFDX = 0,
666 TEXTURE_DFDY = 1,
667 };
668
669 typedef struct
670 __attribute__((__packed__))
671 {
672 unsigned type : 4;
673 unsigned next_type : 4;
674
675 enum mali_texture_op op : 4;
676 unsigned mode : 4;
677
678 /* A little obscure, but last is set for the last texture operation in
679 * a shader. cont appears to just be last's opposite (?). Yeah, I know,
680 * kind of funky.. BiOpen thinks it could do with memory hinting, or
681 * tile locking? */
682
683 unsigned cont : 1;
684 unsigned last : 1;
685
686 unsigned format : 2;
687
688 /* Are sampler_handle/texture_handler respectively set by registers? If
689 * true, the lower 8-bits of the respective field is a register word.
690 * If false, they are an immediate */
691
692 unsigned sampler_register : 1;
693 unsigned texture_register : 1;
694
695 /* Is a register used to specify the
696 * LOD/bias/offset? If set, use the `bias` field as
697 * a register index. If clear, use the `bias` field
698 * as an immediate. */
699 unsigned lod_register : 1;
700
701 /* Is a register used to specify an offset? If set, use the
702 * offset_reg_* fields to encode this, duplicated for each of the
703 * components. If clear, there is implcitly always an immediate offst
704 * specificed in offset_imm_* */
705 unsigned offset_register : 1;
706
707 unsigned in_reg_full : 1;
708 unsigned in_reg_select : 1;
709 unsigned in_reg_upper : 1;
710 unsigned in_reg_swizzle : 8;
711
712 unsigned unknown8 : 2;
713
714 unsigned out_full : 1;
715
716 enum mali_sampler_type sampler_type : 2;
717
718 unsigned out_reg_select : 1;
719 unsigned out_upper : 1;
720
721 unsigned mask : 4;
722
723 /* Intriguingly, textures can take an outmod just like alu ops. Int
724 * outmods are not supported as far as I can tell, so this is only
725 * meaningful for float samplers */
726 midgard_outmod_float outmod : 2;
727
728 unsigned swizzle : 8;
729
730 /* These indicate how many bundles after this texture op may be
731 * executed in parallel with this op. We may execute only ALU and
732 * ld/st in parallel (not other textures), and obviously there cannot
733 * be any dependency (the blob appears to forbid even accessing other
734 * channels of a given texture register). */
735
736 unsigned out_of_order : 2;
737 unsigned unknown4 : 10;
738
739 /* In immediate mode, each offset field is an immediate range [0, 7].
740 *
741 * In register mode, offset_x becomes a register (full, select, upper)
742 * triplet followed by a vec3 swizzle is splattered across
743 * offset_y/offset_z in a genuinely bizarre way.
744 *
745 * For texel fetches in immediate mode, the range is the full [-8, 7],
746 * but for normal texturing the top bit must be zero and a register
747 * used instead. It's not clear where this limitation is from.
748 *
749 * union {
750 * struct {
751 * signed offset_x : 4;
752 * signed offset_y : 4;
753 * signed offset_z : 4;
754 * } immediate;
755 * struct {
756 * bool full : 1;
757 * bool select : 1;
758 * bool upper : 1;
759 * unsigned swizzle : 8;
760 * unsigned zero : 1;
761 * } register;
762 * }
763 */
764
765 unsigned offset : 12;
766
767 /* In immediate bias mode, for a normal texture op, this is
768 * texture bias, computed as int(2^8 * frac(biasf)), with
769 * bias_int = floor(bias). For a textureLod, it's that, but
770 * s/bias/lod. For a texel fetch, this is the LOD as-is.
771 *
772 * In register mode, this is a midgard_tex_register_select
773 * structure and bias_int is zero */
774
775 unsigned bias : 8;
776 signed bias_int : 8;
777
778 /* If sampler/texture_register is set, the bottom 8-bits are
779 * midgard_tex_register_select and the top 8-bits are zero. If they are
780 * clear, they are immediate texture indices */
781
782 unsigned sampler_handle : 16;
783 unsigned texture_handle : 16;
784 }
785 midgard_texture_word;
786
787 /* Technically barriers are texture instructions but it's less work to add them
788 * as an explicitly zeroed special case, since most fields are forced to go to
789 * zero */
790
791 typedef struct
792 __attribute__((__packed__))
793 {
794 unsigned type : 4;
795 unsigned next_type : 4;
796
797 /* op = TEXTURE_OP_BARRIER */
798 unsigned op : 6;
799 unsigned zero1 : 2;
800
801 /* Since helper invocations don't make any sense, these are forced to one */
802 unsigned cont : 1;
803 unsigned last : 1;
804 unsigned zero2 : 14;
805
806 unsigned zero3 : 24;
807 unsigned out_of_order : 4;
808 unsigned zero4 : 4;
809
810 uint64_t zero5;
811 } midgard_texture_barrier_word;
812
813 typedef union midgard_constants {
814 double f64[2];
815 uint64_t u64[2];
816 int64_t i64[2];
817 float f32[4];
818 uint32_t u32[4];
819 int32_t i32[4];
820 uint16_t f16[8];
821 uint16_t u16[8];
822 int16_t i16[8];
823 uint8_t u8[16];
824 int8_t i8[16];
825 }
826 midgard_constants;
827
828 enum midgard_roundmode {
829 MIDGARD_RTE = 0x0, /* round to even */
830 MIDGARD_RTZ = 0x1, /* round to zero */
831 MIDGARD_RTN = 0x2, /* round to negative */
832 MIDGARD_RTP = 0x3, /* round to positive */
833 };
834
835 #endif